]>
Commit | Line | Data |
---|---|---|
dea61d92 | 1 | /* Loop distribution. |
cbe34bb5 | 2 | Copyright (C) 2006-2017 Free Software Foundation, Inc. |
dea61d92 SP |
3 | Contributed by Georges-Andre Silber <Georges-Andre.Silber@ensmp.fr> |
4 | and Sebastian Pop <sebastian.pop@amd.com>. | |
5 | ||
6 | This file is part of GCC. | |
b8698a0f | 7 | |
dea61d92 SP |
8 | GCC is free software; you can redistribute it and/or modify it |
9 | under the terms of the GNU General Public License as published by the | |
10 | Free Software Foundation; either version 3, or (at your option) any | |
11 | later version. | |
b8698a0f | 12 | |
dea61d92 SP |
13 | GCC is distributed in the hope that it will be useful, but WITHOUT |
14 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
b8698a0f | 17 | |
dea61d92 SP |
18 | You should have received a copy of the GNU General Public License |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | /* This pass performs loop distribution: for example, the loop | |
23 | ||
24 | |DO I = 2, N | |
25 | | A(I) = B(I) + C | |
26 | | D(I) = A(I-1)*E | |
27 | |ENDDO | |
28 | ||
b8698a0f | 29 | is transformed to |
dea61d92 SP |
30 | |
31 | |DOALL I = 2, N | |
32 | | A(I) = B(I) + C | |
33 | |ENDDO | |
34 | | | |
35 | |DOALL I = 2, N | |
36 | | D(I) = A(I-1)*E | |
37 | |ENDDO | |
38 | ||
39 | This pass uses an RDG, Reduced Dependence Graph built on top of the | |
40 | data dependence relations. The RDG is then topologically sorted to | |
41 | obtain a map of information producers/consumers based on which it | |
42 | generates the new loops. */ | |
43 | ||
44 | #include "config.h" | |
45 | #include "system.h" | |
46 | #include "coretypes.h" | |
c7131fb2 | 47 | #include "backend.h" |
40e23961 | 48 | #include "tree.h" |
c7131fb2 | 49 | #include "gimple.h" |
957060b5 AM |
50 | #include "cfghooks.h" |
51 | #include "tree-pass.h" | |
c7131fb2 | 52 | #include "ssa.h" |
957060b5 | 53 | #include "gimple-pretty-print.h" |
c7131fb2 | 54 | #include "fold-const.h" |
60393bbc | 55 | #include "cfganal.h" |
5be5c238 | 56 | #include "gimple-iterator.h" |
18f429e2 | 57 | #include "gimplify-me.h" |
d8a2d370 | 58 | #include "stor-layout.h" |
442b4905 | 59 | #include "tree-cfg.h" |
e28030cf | 60 | #include "tree-ssa-loop-manip.h" |
442b4905 AM |
61 | #include "tree-ssa-loop.h" |
62 | #include "tree-into-ssa.h" | |
7a300452 | 63 | #include "tree-ssa.h" |
dea61d92 | 64 | #include "cfgloop.h" |
dea61d92 | 65 | #include "tree-scalar-evolution.h" |
826a536d | 66 | #include "tree-vectorizer.h" |
80ab0b19 RB |
67 | |
68 | ||
69 | /* A Reduced Dependence Graph (RDG) vertex representing a statement. */ | |
526ceb68 | 70 | struct rdg_vertex |
80ab0b19 RB |
71 | { |
72 | /* The statement represented by this vertex. */ | |
355fe088 | 73 | gimple *stmt; |
80ab0b19 RB |
74 | |
75 | /* Vector of data-references in this statement. */ | |
76 | vec<data_reference_p> datarefs; | |
77 | ||
78 | /* True when the statement contains a write to memory. */ | |
79 | bool has_mem_write; | |
80 | ||
81 | /* True when the statement contains a read from memory. */ | |
82 | bool has_mem_reads; | |
526ceb68 | 83 | }; |
80ab0b19 RB |
84 | |
85 | #define RDGV_STMT(V) ((struct rdg_vertex *) ((V)->data))->stmt | |
86 | #define RDGV_DATAREFS(V) ((struct rdg_vertex *) ((V)->data))->datarefs | |
87 | #define RDGV_HAS_MEM_WRITE(V) ((struct rdg_vertex *) ((V)->data))->has_mem_write | |
88 | #define RDGV_HAS_MEM_READS(V) ((struct rdg_vertex *) ((V)->data))->has_mem_reads | |
89 | #define RDG_STMT(RDG, I) RDGV_STMT (&(RDG->vertices[I])) | |
90 | #define RDG_DATAREFS(RDG, I) RDGV_DATAREFS (&(RDG->vertices[I])) | |
91 | #define RDG_MEM_WRITE_STMT(RDG, I) RDGV_HAS_MEM_WRITE (&(RDG->vertices[I])) | |
92 | #define RDG_MEM_READS_STMT(RDG, I) RDGV_HAS_MEM_READS (&(RDG->vertices[I])) | |
93 | ||
94 | /* Data dependence type. */ | |
95 | ||
96 | enum rdg_dep_type | |
97 | { | |
98 | /* Read After Write (RAW). */ | |
99 | flow_dd = 'f', | |
100 | ||
36875e8f RB |
101 | /* Control dependence (execute conditional on). */ |
102 | control_dd = 'c' | |
80ab0b19 RB |
103 | }; |
104 | ||
105 | /* Dependence information attached to an edge of the RDG. */ | |
106 | ||
526ceb68 | 107 | struct rdg_edge |
80ab0b19 RB |
108 | { |
109 | /* Type of the dependence. */ | |
110 | enum rdg_dep_type type; | |
526ceb68 | 111 | }; |
80ab0b19 RB |
112 | |
113 | #define RDGE_TYPE(E) ((struct rdg_edge *) ((E)->data))->type | |
80ab0b19 | 114 | |
80ab0b19 RB |
115 | /* Dump vertex I in RDG to FILE. */ |
116 | ||
117 | static void | |
118 | dump_rdg_vertex (FILE *file, struct graph *rdg, int i) | |
119 | { | |
120 | struct vertex *v = &(rdg->vertices[i]); | |
121 | struct graph_edge *e; | |
122 | ||
123 | fprintf (file, "(vertex %d: (%s%s) (in:", i, | |
124 | RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "", | |
125 | RDG_MEM_READS_STMT (rdg, i) ? "r" : ""); | |
126 | ||
127 | if (v->pred) | |
128 | for (e = v->pred; e; e = e->pred_next) | |
129 | fprintf (file, " %d", e->src); | |
130 | ||
131 | fprintf (file, ") (out:"); | |
132 | ||
133 | if (v->succ) | |
134 | for (e = v->succ; e; e = e->succ_next) | |
135 | fprintf (file, " %d", e->dest); | |
136 | ||
137 | fprintf (file, ")\n"); | |
138 | print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS); | |
139 | fprintf (file, ")\n"); | |
140 | } | |
141 | ||
142 | /* Call dump_rdg_vertex on stderr. */ | |
143 | ||
144 | DEBUG_FUNCTION void | |
145 | debug_rdg_vertex (struct graph *rdg, int i) | |
146 | { | |
147 | dump_rdg_vertex (stderr, rdg, i); | |
148 | } | |
149 | ||
80ab0b19 RB |
150 | /* Dump the reduced dependence graph RDG to FILE. */ |
151 | ||
152 | static void | |
153 | dump_rdg (FILE *file, struct graph *rdg) | |
154 | { | |
80ab0b19 | 155 | fprintf (file, "(rdg\n"); |
2fd5894f RB |
156 | for (int i = 0; i < rdg->n_vertices; i++) |
157 | dump_rdg_vertex (file, rdg, i); | |
80ab0b19 | 158 | fprintf (file, ")\n"); |
80ab0b19 RB |
159 | } |
160 | ||
161 | /* Call dump_rdg on stderr. */ | |
162 | ||
163 | DEBUG_FUNCTION void | |
164 | debug_rdg (struct graph *rdg) | |
165 | { | |
166 | dump_rdg (stderr, rdg); | |
167 | } | |
168 | ||
169 | static void | |
170 | dot_rdg_1 (FILE *file, struct graph *rdg) | |
171 | { | |
172 | int i; | |
174ec470 RB |
173 | pretty_printer buffer; |
174 | pp_needs_newline (&buffer) = false; | |
175 | buffer.buffer->stream = file; | |
80ab0b19 RB |
176 | |
177 | fprintf (file, "digraph RDG {\n"); | |
178 | ||
179 | for (i = 0; i < rdg->n_vertices; i++) | |
180 | { | |
181 | struct vertex *v = &(rdg->vertices[i]); | |
182 | struct graph_edge *e; | |
183 | ||
174ec470 RB |
184 | fprintf (file, "%d [label=\"[%d] ", i, i); |
185 | pp_gimple_stmt_1 (&buffer, RDGV_STMT (v), 0, TDF_SLIM); | |
186 | pp_flush (&buffer); | |
187 | fprintf (file, "\"]\n"); | |
188 | ||
80ab0b19 RB |
189 | /* Highlight reads from memory. */ |
190 | if (RDG_MEM_READS_STMT (rdg, i)) | |
191 | fprintf (file, "%d [style=filled, fillcolor=green]\n", i); | |
192 | ||
193 | /* Highlight stores to memory. */ | |
194 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
195 | fprintf (file, "%d [style=filled, fillcolor=red]\n", i); | |
196 | ||
197 | if (v->succ) | |
198 | for (e = v->succ; e; e = e->succ_next) | |
199 | switch (RDGE_TYPE (e)) | |
200 | { | |
80ab0b19 RB |
201 | case flow_dd: |
202 | /* These are the most common dependences: don't print these. */ | |
203 | fprintf (file, "%d -> %d \n", i, e->dest); | |
204 | break; | |
205 | ||
36875e8f RB |
206 | case control_dd: |
207 | fprintf (file, "%d -> %d [label=control] \n", i, e->dest); | |
208 | break; | |
209 | ||
80ab0b19 RB |
210 | default: |
211 | gcc_unreachable (); | |
212 | } | |
213 | } | |
214 | ||
215 | fprintf (file, "}\n\n"); | |
216 | } | |
217 | ||
218 | /* Display the Reduced Dependence Graph using dotty. */ | |
219 | ||
220 | DEBUG_FUNCTION void | |
221 | dot_rdg (struct graph *rdg) | |
222 | { | |
174ec470 | 223 | /* When debugging, you may want to enable the following code. */ |
b6d94045 | 224 | #ifdef HAVE_POPEN |
c3284718 | 225 | FILE *file = popen ("dot -Tx11", "w"); |
174ec470 RB |
226 | if (!file) |
227 | return; | |
80ab0b19 | 228 | dot_rdg_1 (file, rdg); |
174ec470 RB |
229 | fflush (file); |
230 | close (fileno (file)); | |
231 | pclose (file); | |
80ab0b19 RB |
232 | #else |
233 | dot_rdg_1 (stderr, rdg); | |
234 | #endif | |
235 | } | |
236 | ||
237 | /* Returns the index of STMT in RDG. */ | |
238 | ||
239 | static int | |
355fe088 | 240 | rdg_vertex_for_stmt (struct graph *rdg ATTRIBUTE_UNUSED, gimple *stmt) |
80ab0b19 RB |
241 | { |
242 | int index = gimple_uid (stmt); | |
243 | gcc_checking_assert (index == -1 || RDG_STMT (rdg, index) == stmt); | |
244 | return index; | |
245 | } | |
246 | ||
80ab0b19 RB |
247 | /* Creates dependence edges in RDG for all the uses of DEF. IDEF is |
248 | the index of DEF in RDG. */ | |
249 | ||
250 | static void | |
251 | create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef) | |
252 | { | |
253 | use_operand_p imm_use_p; | |
254 | imm_use_iterator iterator; | |
255 | ||
256 | FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def) | |
257 | { | |
258 | struct graph_edge *e; | |
259 | int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p)); | |
260 | ||
261 | if (use < 0) | |
262 | continue; | |
263 | ||
264 | e = add_edge (rdg, idef, use); | |
265 | e->data = XNEW (struct rdg_edge); | |
266 | RDGE_TYPE (e) = flow_dd; | |
80ab0b19 RB |
267 | } |
268 | } | |
269 | ||
36875e8f RB |
270 | /* Creates an edge for the control dependences of BB to the vertex V. */ |
271 | ||
272 | static void | |
273 | create_edge_for_control_dependence (struct graph *rdg, basic_block bb, | |
274 | int v, control_dependences *cd) | |
275 | { | |
276 | bitmap_iterator bi; | |
277 | unsigned edge_n; | |
278 | EXECUTE_IF_SET_IN_BITMAP (cd->get_edges_dependent_on (bb->index), | |
279 | 0, edge_n, bi) | |
280 | { | |
30fd2977 | 281 | basic_block cond_bb = cd->get_edge_src (edge_n); |
355fe088 | 282 | gimple *stmt = last_stmt (cond_bb); |
36875e8f RB |
283 | if (stmt && is_ctrl_stmt (stmt)) |
284 | { | |
285 | struct graph_edge *e; | |
286 | int c = rdg_vertex_for_stmt (rdg, stmt); | |
287 | if (c < 0) | |
288 | continue; | |
289 | ||
290 | e = add_edge (rdg, c, v); | |
291 | e->data = XNEW (struct rdg_edge); | |
292 | RDGE_TYPE (e) = control_dd; | |
36875e8f RB |
293 | } |
294 | } | |
295 | } | |
296 | ||
80ab0b19 RB |
297 | /* Creates the edges of the reduced dependence graph RDG. */ |
298 | ||
299 | static void | |
447f3223 | 300 | create_rdg_flow_edges (struct graph *rdg) |
80ab0b19 RB |
301 | { |
302 | int i; | |
80ab0b19 RB |
303 | def_operand_p def_p; |
304 | ssa_op_iter iter; | |
305 | ||
80ab0b19 RB |
306 | for (i = 0; i < rdg->n_vertices; i++) |
307 | FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i), | |
308 | iter, SSA_OP_DEF) | |
309 | create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i); | |
447f3223 | 310 | } |
36875e8f | 311 | |
447f3223 RB |
312 | /* Creates the edges of the reduced dependence graph RDG. */ |
313 | ||
314 | static void | |
b71b7a8e | 315 | create_rdg_cd_edges (struct graph *rdg, control_dependences *cd, loop_p loop) |
447f3223 RB |
316 | { |
317 | int i; | |
318 | ||
319 | for (i = 0; i < rdg->n_vertices; i++) | |
320 | { | |
355fe088 | 321 | gimple *stmt = RDG_STMT (rdg, i); |
447f3223 RB |
322 | if (gimple_code (stmt) == GIMPLE_PHI) |
323 | { | |
324 | edge_iterator ei; | |
325 | edge e; | |
326 | FOR_EACH_EDGE (e, ei, gimple_bb (stmt)->preds) | |
b71b7a8e | 327 | if (flow_bb_inside_loop_p (loop, e->src)) |
36875e8f | 328 | create_edge_for_control_dependence (rdg, e->src, i, cd); |
447f3223 RB |
329 | } |
330 | else | |
331 | create_edge_for_control_dependence (rdg, gimple_bb (stmt), i, cd); | |
332 | } | |
80ab0b19 RB |
333 | } |
334 | ||
335 | /* Build the vertices of the reduced dependence graph RDG. Return false | |
336 | if that failed. */ | |
337 | ||
338 | static bool | |
355fe088 | 339 | create_rdg_vertices (struct graph *rdg, vec<gimple *> stmts, loop_p loop, |
80ab0b19 RB |
340 | vec<data_reference_p> *datarefs) |
341 | { | |
342 | int i; | |
355fe088 | 343 | gimple *stmt; |
80ab0b19 RB |
344 | |
345 | FOR_EACH_VEC_ELT (stmts, i, stmt) | |
346 | { | |
347 | struct vertex *v = &(rdg->vertices[i]); | |
348 | ||
349 | /* Record statement to vertex mapping. */ | |
350 | gimple_set_uid (stmt, i); | |
351 | ||
352 | v->data = XNEW (struct rdg_vertex); | |
353 | RDGV_STMT (v) = stmt; | |
354 | RDGV_DATAREFS (v).create (0); | |
355 | RDGV_HAS_MEM_WRITE (v) = false; | |
356 | RDGV_HAS_MEM_READS (v) = false; | |
357 | if (gimple_code (stmt) == GIMPLE_PHI) | |
358 | continue; | |
359 | ||
360 | unsigned drp = datarefs->length (); | |
361 | if (!find_data_references_in_stmt (loop, stmt, datarefs)) | |
362 | return false; | |
363 | for (unsigned j = drp; j < datarefs->length (); ++j) | |
364 | { | |
365 | data_reference_p dr = (*datarefs)[j]; | |
366 | if (DR_IS_READ (dr)) | |
367 | RDGV_HAS_MEM_READS (v) = true; | |
368 | else | |
369 | RDGV_HAS_MEM_WRITE (v) = true; | |
370 | RDGV_DATAREFS (v).safe_push (dr); | |
371 | } | |
372 | } | |
373 | return true; | |
374 | } | |
375 | ||
2fd5894f | 376 | /* Initialize STMTS with all the statements of LOOP. The order in |
80ab0b19 RB |
377 | which we discover statements is important as |
378 | generate_loops_for_partition is using the same traversal for | |
2fd5894f | 379 | identifying statements in loop copies. */ |
80ab0b19 RB |
380 | |
381 | static void | |
355fe088 | 382 | stmts_from_loop (struct loop *loop, vec<gimple *> *stmts) |
80ab0b19 RB |
383 | { |
384 | unsigned int i; | |
385 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
386 | ||
387 | for (i = 0; i < loop->num_nodes; i++) | |
388 | { | |
389 | basic_block bb = bbs[i]; | |
80ab0b19 | 390 | |
538dd0b7 DM |
391 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
392 | gsi_next (&bsi)) | |
393 | if (!virtual_operand_p (gimple_phi_result (bsi.phi ()))) | |
394 | stmts->safe_push (bsi.phi ()); | |
80ab0b19 | 395 | |
538dd0b7 DM |
396 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
397 | gsi_next (&bsi)) | |
80ab0b19 | 398 | { |
355fe088 | 399 | gimple *stmt = gsi_stmt (bsi); |
80ab0b19 RB |
400 | if (gimple_code (stmt) != GIMPLE_LABEL && !is_gimple_debug (stmt)) |
401 | stmts->safe_push (stmt); | |
402 | } | |
403 | } | |
404 | ||
405 | free (bbs); | |
406 | } | |
407 | ||
80ab0b19 RB |
408 | /* Free the reduced dependence graph RDG. */ |
409 | ||
410 | static void | |
411 | free_rdg (struct graph *rdg) | |
412 | { | |
413 | int i; | |
414 | ||
415 | for (i = 0; i < rdg->n_vertices; i++) | |
416 | { | |
417 | struct vertex *v = &(rdg->vertices[i]); | |
418 | struct graph_edge *e; | |
419 | ||
420 | for (e = v->succ; e; e = e->succ_next) | |
447f3223 | 421 | free (e->data); |
80ab0b19 RB |
422 | |
423 | if (v->data) | |
424 | { | |
425 | gimple_set_uid (RDGV_STMT (v), -1); | |
426 | free_data_refs (RDGV_DATAREFS (v)); | |
427 | free (v->data); | |
428 | } | |
429 | } | |
430 | ||
431 | free_graph (rdg); | |
432 | } | |
433 | ||
434 | /* Build the Reduced Dependence Graph (RDG) with one vertex per | |
97463b2b | 435 | statement of the loop nest LOOP_NEST, and one edge per data dependence or |
80ab0b19 RB |
436 | scalar dependence. */ |
437 | ||
438 | static struct graph * | |
36875e8f | 439 | build_rdg (vec<loop_p> loop_nest, control_dependences *cd) |
80ab0b19 RB |
440 | { |
441 | struct graph *rdg; | |
80ab0b19 | 442 | vec<data_reference_p> datarefs; |
80ab0b19 | 443 | |
97463b2b | 444 | /* Create the RDG vertices from the stmts of the loop nest. */ |
355fe088 | 445 | auto_vec<gimple *, 10> stmts; |
97463b2b | 446 | stmts_from_loop (loop_nest[0], &stmts); |
24f161fd | 447 | rdg = new_graph (stmts.length ()); |
80ab0b19 | 448 | datarefs.create (10); |
97463b2b | 449 | if (!create_rdg_vertices (rdg, stmts, loop_nest[0], &datarefs)) |
80ab0b19 | 450 | { |
97463b2b | 451 | datarefs.release (); |
80ab0b19 RB |
452 | free_rdg (rdg); |
453 | return NULL; | |
454 | } | |
455 | stmts.release (); | |
97463b2b | 456 | |
447f3223 RB |
457 | create_rdg_flow_edges (rdg); |
458 | if (cd) | |
b71b7a8e | 459 | create_rdg_cd_edges (rdg, cd, loop_nest[0]); |
447f3223 | 460 | |
97463b2b | 461 | datarefs.release (); |
80ab0b19 RB |
462 | |
463 | return rdg; | |
464 | } | |
465 | ||
80ab0b19 | 466 | |
dea61d92 | 467 | |
b9fc0497 | 468 | enum partition_kind { |
510d73a0 | 469 | PKIND_NORMAL, PKIND_MEMSET, PKIND_MEMCPY, PKIND_MEMMOVE |
b9fc0497 | 470 | }; |
30d55936 | 471 | |
526ceb68 | 472 | struct partition |
c61f8985 RG |
473 | { |
474 | bitmap stmts; | |
80ab0b19 | 475 | bitmap loops; |
826a536d | 476 | bool reduction_p; |
34e82342 | 477 | bool plus_one; |
30d55936 | 478 | enum partition_kind kind; |
d0582dc1 RG |
479 | /* data-references a kind != PKIND_NORMAL partition is about. */ |
480 | data_reference_p main_dr; | |
481 | data_reference_p secondary_dr; | |
818625cf | 482 | tree niter; |
526ceb68 | 483 | }; |
c61f8985 | 484 | |
c61f8985 RG |
485 | |
486 | /* Allocate and initialize a partition from BITMAP. */ | |
487 | ||
526ceb68 | 488 | static partition * |
80ab0b19 | 489 | partition_alloc (bitmap stmts, bitmap loops) |
c61f8985 | 490 | { |
526ceb68 | 491 | partition *partition = XCNEW (struct partition); |
c61f8985 | 492 | partition->stmts = stmts ? stmts : BITMAP_ALLOC (NULL); |
80ab0b19 | 493 | partition->loops = loops ? loops : BITMAP_ALLOC (NULL); |
826a536d | 494 | partition->reduction_p = false; |
30d55936 | 495 | partition->kind = PKIND_NORMAL; |
c61f8985 RG |
496 | return partition; |
497 | } | |
498 | ||
499 | /* Free PARTITION. */ | |
500 | ||
501 | static void | |
526ceb68 | 502 | partition_free (partition *partition) |
c61f8985 RG |
503 | { |
504 | BITMAP_FREE (partition->stmts); | |
80ab0b19 | 505 | BITMAP_FREE (partition->loops); |
c61f8985 RG |
506 | free (partition); |
507 | } | |
508 | ||
30d55936 RG |
509 | /* Returns true if the partition can be generated as a builtin. */ |
510 | ||
511 | static bool | |
526ceb68 | 512 | partition_builtin_p (partition *partition) |
30d55936 | 513 | { |
826a536d | 514 | return partition->kind != PKIND_NORMAL; |
30d55936 | 515 | } |
c61f8985 | 516 | |
826a536d | 517 | /* Returns true if the partition contains a reduction. */ |
7ad672e4 RG |
518 | |
519 | static bool | |
526ceb68 | 520 | partition_reduction_p (partition *partition) |
7ad672e4 | 521 | { |
826a536d | 522 | return partition->reduction_p; |
7ad672e4 RG |
523 | } |
524 | ||
826a536d RB |
525 | /* Merge PARTITION into the partition DEST. */ |
526 | ||
527 | static void | |
526ceb68 | 528 | partition_merge_into (partition *dest, partition *partition) |
826a536d RB |
529 | { |
530 | dest->kind = PKIND_NORMAL; | |
531 | bitmap_ior_into (dest->stmts, partition->stmts); | |
532 | if (partition_reduction_p (partition)) | |
533 | dest->reduction_p = true; | |
534 | } | |
535 | ||
536 | ||
c07a8cb3 RG |
537 | /* Returns true when DEF is an SSA_NAME defined in LOOP and used after |
538 | the LOOP. */ | |
539 | ||
540 | static bool | |
541 | ssa_name_has_uses_outside_loop_p (tree def, loop_p loop) | |
542 | { | |
543 | imm_use_iterator imm_iter; | |
544 | use_operand_p use_p; | |
545 | ||
546 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
e665269a | 547 | { |
355fe088 | 548 | gimple *use_stmt = USE_STMT (use_p); |
e665269a RG |
549 | if (!is_gimple_debug (use_stmt) |
550 | && loop != loop_containing_stmt (use_stmt)) | |
551 | return true; | |
552 | } | |
c07a8cb3 RG |
553 | |
554 | return false; | |
555 | } | |
556 | ||
557 | /* Returns true when STMT defines a scalar variable used after the | |
88af7c1a | 558 | loop LOOP. */ |
c07a8cb3 RG |
559 | |
560 | static bool | |
355fe088 | 561 | stmt_has_scalar_dependences_outside_loop (loop_p loop, gimple *stmt) |
c07a8cb3 | 562 | { |
88af7c1a RG |
563 | def_operand_p def_p; |
564 | ssa_op_iter op_iter; | |
c07a8cb3 | 565 | |
9ca86fc3 RG |
566 | if (gimple_code (stmt) == GIMPLE_PHI) |
567 | return ssa_name_has_uses_outside_loop_p (gimple_phi_result (stmt), loop); | |
568 | ||
88af7c1a RG |
569 | FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_DEF) |
570 | if (ssa_name_has_uses_outside_loop_p (DEF_FROM_PTR (def_p), loop)) | |
571 | return true; | |
c07a8cb3 | 572 | |
88af7c1a | 573 | return false; |
c07a8cb3 RG |
574 | } |
575 | ||
dea61d92 SP |
576 | /* Return a copy of LOOP placed before LOOP. */ |
577 | ||
578 | static struct loop * | |
579 | copy_loop_before (struct loop *loop) | |
580 | { | |
581 | struct loop *res; | |
582 | edge preheader = loop_preheader_edge (loop); | |
583 | ||
dea61d92 | 584 | initialize_original_copy_tables (); |
5ce9450f | 585 | res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, NULL, preheader); |
30d55936 | 586 | gcc_assert (res != NULL); |
dea61d92 | 587 | free_original_copy_tables (); |
2cfc56b9 | 588 | delete_update_ssa (); |
dea61d92 SP |
589 | |
590 | return res; | |
591 | } | |
592 | ||
593 | /* Creates an empty basic block after LOOP. */ | |
594 | ||
595 | static void | |
596 | create_bb_after_loop (struct loop *loop) | |
597 | { | |
598 | edge exit = single_exit (loop); | |
599 | ||
600 | if (!exit) | |
601 | return; | |
602 | ||
603 | split_edge (exit); | |
604 | } | |
605 | ||
606 | /* Generate code for PARTITION from the code in LOOP. The loop is | |
607 | copied when COPY_P is true. All the statements not flagged in the | |
608 | PARTITION bitmap are removed from the loop or from its copy. The | |
609 | statements are indexed in sequence inside a basic block, and the | |
30d55936 | 610 | basic blocks of a loop are taken in dom order. */ |
dea61d92 | 611 | |
30d55936 | 612 | static void |
526ceb68 | 613 | generate_loops_for_partition (struct loop *loop, partition *partition, |
c61f8985 | 614 | bool copy_p) |
dea61d92 | 615 | { |
2fd5894f | 616 | unsigned i; |
dea61d92 SP |
617 | basic_block *bbs; |
618 | ||
619 | if (copy_p) | |
620 | { | |
621 | loop = copy_loop_before (loop); | |
30d55936 | 622 | gcc_assert (loop != NULL); |
dea61d92 SP |
623 | create_preheader (loop, CP_SIMPLE_PREHEADERS); |
624 | create_bb_after_loop (loop); | |
625 | } | |
626 | ||
2fd5894f | 627 | /* Remove stmts not in the PARTITION bitmap. */ |
dea61d92 SP |
628 | bbs = get_loop_body_in_dom_order (loop); |
629 | ||
b03c3082 | 630 | if (MAY_HAVE_DEBUG_STMTS) |
2fd5894f | 631 | for (i = 0; i < loop->num_nodes; i++) |
b03c3082 JJ |
632 | { |
633 | basic_block bb = bbs[i]; | |
634 | ||
538dd0b7 DM |
635 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
636 | gsi_next (&bsi)) | |
2fd5894f | 637 | { |
538dd0b7 | 638 | gphi *phi = bsi.phi (); |
2fd5894f RB |
639 | if (!virtual_operand_p (gimple_phi_result (phi)) |
640 | && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
641 | reset_debug_uses (phi); | |
642 | } | |
b03c3082 | 643 | |
538dd0b7 | 644 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
b03c3082 | 645 | { |
355fe088 | 646 | gimple *stmt = gsi_stmt (bsi); |
b03c3082 JJ |
647 | if (gimple_code (stmt) != GIMPLE_LABEL |
648 | && !is_gimple_debug (stmt) | |
2fd5894f | 649 | && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) |
b03c3082 JJ |
650 | reset_debug_uses (stmt); |
651 | } | |
652 | } | |
653 | ||
2fd5894f | 654 | for (i = 0; i < loop->num_nodes; i++) |
dea61d92 SP |
655 | { |
656 | basic_block bb = bbs[i]; | |
dea61d92 | 657 | |
538dd0b7 | 658 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);) |
2fd5894f | 659 | { |
538dd0b7 | 660 | gphi *phi = bsi.phi (); |
2fd5894f RB |
661 | if (!virtual_operand_p (gimple_phi_result (phi)) |
662 | && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
2706a615 | 663 | remove_phi_node (&bsi, true); |
2fd5894f RB |
664 | else |
665 | gsi_next (&bsi); | |
666 | } | |
dea61d92 | 667 | |
538dd0b7 | 668 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);) |
2706a615 | 669 | { |
355fe088 | 670 | gimple *stmt = gsi_stmt (bsi); |
b03c3082 JJ |
671 | if (gimple_code (stmt) != GIMPLE_LABEL |
672 | && !is_gimple_debug (stmt) | |
2fd5894f | 673 | && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) |
2706a615 | 674 | { |
36875e8f RB |
675 | /* Choose an arbitrary path through the empty CFG part |
676 | that this unnecessary control stmt controls. */ | |
538dd0b7 | 677 | if (gcond *cond_stmt = dyn_cast <gcond *> (stmt)) |
36875e8f | 678 | { |
538dd0b7 | 679 | gimple_cond_make_false (cond_stmt); |
36875e8f RB |
680 | update_stmt (stmt); |
681 | } | |
682 | else if (gimple_code (stmt) == GIMPLE_SWITCH) | |
683 | { | |
538dd0b7 | 684 | gswitch *switch_stmt = as_a <gswitch *> (stmt); |
36875e8f | 685 | gimple_switch_set_index |
538dd0b7 | 686 | (switch_stmt, CASE_LOW (gimple_switch_label (switch_stmt, 1))); |
36875e8f RB |
687 | update_stmt (stmt); |
688 | } | |
689 | else | |
690 | { | |
691 | unlink_stmt_vdef (stmt); | |
692 | gsi_remove (&bsi, true); | |
693 | release_defs (stmt); | |
694 | continue; | |
695 | } | |
2706a615 | 696 | } |
36875e8f | 697 | gsi_next (&bsi); |
2706a615 | 698 | } |
dea61d92 SP |
699 | } |
700 | ||
701 | free (bbs); | |
dea61d92 SP |
702 | } |
703 | ||
d0582dc1 | 704 | /* Build the size argument for a memory operation call. */ |
3661e899 | 705 | |
d0582dc1 | 706 | static tree |
d995e887 RB |
707 | build_size_arg_loc (location_t loc, data_reference_p dr, tree nb_iter, |
708 | bool plus_one) | |
3661e899 | 709 | { |
d995e887 RB |
710 | tree size = fold_convert_loc (loc, sizetype, nb_iter); |
711 | if (plus_one) | |
712 | size = size_binop (PLUS_EXPR, size, size_one_node); | |
713 | size = fold_build2_loc (loc, MULT_EXPR, sizetype, size, | |
d0582dc1 | 714 | TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr)))); |
d995e887 RB |
715 | size = fold_convert_loc (loc, size_type_node, size); |
716 | return size; | |
d0582dc1 RG |
717 | } |
718 | ||
719 | /* Build an address argument for a memory operation call. */ | |
720 | ||
721 | static tree | |
722 | build_addr_arg_loc (location_t loc, data_reference_p dr, tree nb_bytes) | |
723 | { | |
724 | tree addr_base; | |
725 | ||
726 | addr_base = size_binop_loc (loc, PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr)); | |
727 | addr_base = fold_convert_loc (loc, sizetype, addr_base); | |
728 | ||
729 | /* Test for a negative stride, iterating over every element. */ | |
730 | if (tree_int_cst_sgn (DR_STEP (dr)) == -1) | |
731 | { | |
732 | addr_base = size_binop_loc (loc, MINUS_EXPR, addr_base, | |
733 | fold_convert_loc (loc, sizetype, nb_bytes)); | |
734 | addr_base = size_binop_loc (loc, PLUS_EXPR, addr_base, | |
735 | TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr)))); | |
736 | } | |
737 | ||
738 | return fold_build_pointer_plus_loc (loc, DR_BASE_ADDRESS (dr), addr_base); | |
3661e899 TB |
739 | } |
740 | ||
401f3a81 JJ |
741 | /* If VAL memory representation contains the same value in all bytes, |
742 | return that value, otherwise return -1. | |
743 | E.g. for 0x24242424 return 0x24, for IEEE double | |
744 | 747708026454360457216.0 return 0x44, etc. */ | |
745 | ||
746 | static int | |
747 | const_with_all_bytes_same (tree val) | |
748 | { | |
749 | unsigned char buf[64]; | |
750 | int i, len; | |
751 | ||
752 | if (integer_zerop (val) | |
401f3a81 JJ |
753 | || (TREE_CODE (val) == CONSTRUCTOR |
754 | && !TREE_CLOBBER_P (val) | |
755 | && CONSTRUCTOR_NELTS (val) == 0)) | |
756 | return 0; | |
757 | ||
9e207d6f JJ |
758 | if (real_zerop (val)) |
759 | { | |
760 | /* Only return 0 for +0.0, not for -0.0, which doesn't have | |
761 | an all bytes same memory representation. Don't transform | |
762 | -0.0 stores into +0.0 even for !HONOR_SIGNED_ZEROS. */ | |
763 | switch (TREE_CODE (val)) | |
764 | { | |
765 | case REAL_CST: | |
766 | if (!real_isneg (TREE_REAL_CST_PTR (val))) | |
767 | return 0; | |
768 | break; | |
769 | case COMPLEX_CST: | |
770 | if (!const_with_all_bytes_same (TREE_REALPART (val)) | |
771 | && !const_with_all_bytes_same (TREE_IMAGPART (val))) | |
772 | return 0; | |
773 | break; | |
774 | case VECTOR_CST: | |
775 | unsigned int j; | |
776 | for (j = 0; j < VECTOR_CST_NELTS (val); ++j) | |
66088065 | 777 | if (const_with_all_bytes_same (VECTOR_CST_ELT (val, j))) |
9e207d6f JJ |
778 | break; |
779 | if (j == VECTOR_CST_NELTS (val)) | |
780 | return 0; | |
781 | break; | |
782 | default: | |
783 | break; | |
784 | } | |
785 | } | |
786 | ||
401f3a81 JJ |
787 | if (CHAR_BIT != 8 || BITS_PER_UNIT != 8) |
788 | return -1; | |
789 | ||
790 | len = native_encode_expr (val, buf, sizeof (buf)); | |
791 | if (len == 0) | |
792 | return -1; | |
793 | for (i = 1; i < len; i++) | |
794 | if (buf[i] != buf[0]) | |
795 | return -1; | |
796 | return buf[0]; | |
797 | } | |
798 | ||
30d55936 | 799 | /* Generate a call to memset for PARTITION in LOOP. */ |
dea61d92 | 800 | |
cfee318d | 801 | static void |
526ceb68 | 802 | generate_memset_builtin (struct loop *loop, partition *partition) |
dea61d92 | 803 | { |
30d55936 | 804 | gimple_stmt_iterator gsi; |
355fe088 | 805 | gimple *stmt, *fn_call; |
818625cf | 806 | tree mem, fn, nb_bytes; |
30d55936 | 807 | location_t loc; |
b6dd5261 | 808 | tree val; |
30d55936 | 809 | |
d0582dc1 | 810 | stmt = DR_STMT (partition->main_dr); |
30d55936 | 811 | loc = gimple_location (stmt); |
30d55936 RG |
812 | |
813 | /* The new statements will be placed before LOOP. */ | |
814 | gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
dea61d92 | 815 | |
d995e887 RB |
816 | nb_bytes = build_size_arg_loc (loc, partition->main_dr, partition->niter, |
817 | partition->plus_one); | |
d0582dc1 RG |
818 | nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
819 | false, GSI_CONTINUE_LINKING); | |
820 | mem = build_addr_arg_loc (loc, partition->main_dr, nb_bytes); | |
821 | mem = force_gimple_operand_gsi (&gsi, mem, true, NULL_TREE, | |
822 | false, GSI_CONTINUE_LINKING); | |
dea61d92 | 823 | |
b6dd5261 RG |
824 | /* This exactly matches the pattern recognition in classify_partition. */ |
825 | val = gimple_assign_rhs1 (stmt); | |
401f3a81 JJ |
826 | /* Handle constants like 0x15151515 and similarly |
827 | floating point constants etc. where all bytes are the same. */ | |
828 | int bytev = const_with_all_bytes_same (val); | |
829 | if (bytev != -1) | |
830 | val = build_int_cst (integer_type_node, bytev); | |
831 | else if (TREE_CODE (val) == INTEGER_CST) | |
832 | val = fold_convert (integer_type_node, val); | |
833 | else if (!useless_type_conversion_p (integer_type_node, TREE_TYPE (val))) | |
b6dd5261 | 834 | { |
b731b390 | 835 | tree tem = make_ssa_name (integer_type_node); |
355fe088 | 836 | gimple *cstmt = gimple_build_assign (tem, NOP_EXPR, val); |
401f3a81 JJ |
837 | gsi_insert_after (&gsi, cstmt, GSI_CONTINUE_LINKING); |
838 | val = tem; | |
b6dd5261 RG |
839 | } |
840 | ||
e79983f4 | 841 | fn = build_fold_addr_expr (builtin_decl_implicit (BUILT_IN_MEMSET)); |
b6dd5261 | 842 | fn_call = gimple_build_call (fn, 3, mem, val, nb_bytes); |
d0582dc1 | 843 | gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); |
dea61d92 SP |
844 | |
845 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
b6dd5261 RG |
846 | { |
847 | fprintf (dump_file, "generated memset"); | |
401f3a81 | 848 | if (bytev == 0) |
b6dd5261 | 849 | fprintf (dump_file, " zero\n"); |
b6dd5261 RG |
850 | else |
851 | fprintf (dump_file, "\n"); | |
852 | } | |
dea61d92 SP |
853 | } |
854 | ||
d0582dc1 RG |
855 | /* Generate a call to memcpy for PARTITION in LOOP. */ |
856 | ||
857 | static void | |
526ceb68 | 858 | generate_memcpy_builtin (struct loop *loop, partition *partition) |
d0582dc1 RG |
859 | { |
860 | gimple_stmt_iterator gsi; | |
355fe088 | 861 | gimple *stmt, *fn_call; |
818625cf | 862 | tree dest, src, fn, nb_bytes; |
d0582dc1 RG |
863 | location_t loc; |
864 | enum built_in_function kind; | |
865 | ||
866 | stmt = DR_STMT (partition->main_dr); | |
867 | loc = gimple_location (stmt); | |
d0582dc1 RG |
868 | |
869 | /* The new statements will be placed before LOOP. */ | |
870 | gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
871 | ||
d995e887 RB |
872 | nb_bytes = build_size_arg_loc (loc, partition->main_dr, partition->niter, |
873 | partition->plus_one); | |
d0582dc1 RG |
874 | nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
875 | false, GSI_CONTINUE_LINKING); | |
876 | dest = build_addr_arg_loc (loc, partition->main_dr, nb_bytes); | |
877 | src = build_addr_arg_loc (loc, partition->secondary_dr, nb_bytes); | |
510d73a0 RB |
878 | if (partition->kind == PKIND_MEMCPY |
879 | || ! ptr_derefs_may_alias_p (dest, src)) | |
d0582dc1 | 880 | kind = BUILT_IN_MEMCPY; |
510d73a0 RB |
881 | else |
882 | kind = BUILT_IN_MEMMOVE; | |
d0582dc1 RG |
883 | |
884 | dest = force_gimple_operand_gsi (&gsi, dest, true, NULL_TREE, | |
885 | false, GSI_CONTINUE_LINKING); | |
886 | src = force_gimple_operand_gsi (&gsi, src, true, NULL_TREE, | |
887 | false, GSI_CONTINUE_LINKING); | |
888 | fn = build_fold_addr_expr (builtin_decl_implicit (kind)); | |
889 | fn_call = gimple_build_call (fn, 3, dest, src, nb_bytes); | |
890 | gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); | |
891 | ||
892 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
893 | { | |
894 | if (kind == BUILT_IN_MEMCPY) | |
895 | fprintf (dump_file, "generated memcpy\n"); | |
896 | else | |
897 | fprintf (dump_file, "generated memmove\n"); | |
898 | } | |
899 | } | |
900 | ||
30d55936 | 901 | /* Remove and destroy the loop LOOP. */ |
dea61d92 | 902 | |
30d55936 RG |
903 | static void |
904 | destroy_loop (struct loop *loop) | |
dea61d92 | 905 | { |
30d55936 RG |
906 | unsigned nbbs = loop->num_nodes; |
907 | edge exit = single_exit (loop); | |
908 | basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest; | |
dea61d92 | 909 | basic_block *bbs; |
30d55936 | 910 | unsigned i; |
dea61d92 SP |
911 | |
912 | bbs = get_loop_body_in_dom_order (loop); | |
913 | ||
30d55936 RG |
914 | redirect_edge_pred (exit, src); |
915 | exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE); | |
916 | exit->flags |= EDGE_FALLTHRU; | |
917 | cancel_loop_tree (loop); | |
918 | rescan_loop_exit (exit, false, true); | |
dea61d92 | 919 | |
b9aba0a0 RB |
920 | i = nbbs; |
921 | do | |
c014f6f5 RG |
922 | { |
923 | /* We have made sure to not leave any dangling uses of SSA | |
924 | names defined in the loop. With the exception of virtuals. | |
925 | Make sure we replace all uses of virtual defs that will remain | |
926 | outside of the loop with the bare symbol as delete_basic_block | |
927 | will release them. */ | |
b9aba0a0 | 928 | --i; |
538dd0b7 DM |
929 | for (gphi_iterator gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); |
930 | gsi_next (&gsi)) | |
c014f6f5 | 931 | { |
538dd0b7 | 932 | gphi *phi = gsi.phi (); |
ea057359 | 933 | if (virtual_operand_p (gimple_phi_result (phi))) |
c014f6f5 RG |
934 | mark_virtual_phi_result_for_renaming (phi); |
935 | } | |
538dd0b7 DM |
936 | for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi); |
937 | gsi_next (&gsi)) | |
c014f6f5 | 938 | { |
355fe088 | 939 | gimple *stmt = gsi_stmt (gsi); |
c014f6f5 RG |
940 | tree vdef = gimple_vdef (stmt); |
941 | if (vdef && TREE_CODE (vdef) == SSA_NAME) | |
942 | mark_virtual_operand_for_renaming (vdef); | |
943 | } | |
944 | delete_basic_block (bbs[i]); | |
945 | } | |
b9aba0a0 RB |
946 | while (i != 0); |
947 | ||
dea61d92 | 948 | free (bbs); |
30d55936 RG |
949 | |
950 | set_immediate_dominator (CDI_DOMINATORS, dest, | |
951 | recompute_dominator (CDI_DOMINATORS, dest)); | |
dea61d92 SP |
952 | } |
953 | ||
b71b7a8e | 954 | /* Generates code for PARTITION. Return whether LOOP needs to be destroyed. */ |
dea61d92 | 955 | |
b71b7a8e | 956 | static bool |
d0582dc1 | 957 | generate_code_for_partition (struct loop *loop, |
526ceb68 | 958 | partition *partition, bool copy_p) |
dea61d92 | 959 | { |
30d55936 RG |
960 | switch (partition->kind) |
961 | { | |
826a536d RB |
962 | case PKIND_NORMAL: |
963 | /* Reductions all have to be in the last partition. */ | |
964 | gcc_assert (!partition_reduction_p (partition) | |
965 | || !copy_p); | |
966 | generate_loops_for_partition (loop, partition, copy_p); | |
b71b7a8e | 967 | return false; |
826a536d | 968 | |
30d55936 | 969 | case PKIND_MEMSET: |
d0582dc1 | 970 | generate_memset_builtin (loop, partition); |
d0582dc1 RG |
971 | break; |
972 | ||
973 | case PKIND_MEMCPY: | |
510d73a0 | 974 | case PKIND_MEMMOVE: |
d0582dc1 | 975 | generate_memcpy_builtin (loop, partition); |
30d55936 RG |
976 | break; |
977 | ||
978 | default: | |
979 | gcc_unreachable (); | |
980 | } | |
dea61d92 | 981 | |
826a536d RB |
982 | /* Common tail for partitions we turn into a call. If this was the last |
983 | partition for which we generate code, we have to destroy the loop. */ | |
984 | if (!copy_p) | |
b71b7a8e RB |
985 | return true; |
986 | return false; | |
dea61d92 SP |
987 | } |
988 | ||
dea61d92 | 989 | |
24f161fd RB |
990 | /* Returns a partition with all the statements needed for computing |
991 | the vertex V of the RDG, also including the loop exit conditions. */ | |
dea61d92 | 992 | |
526ceb68 | 993 | static partition * |
24f161fd | 994 | build_rdg_partition_for_vertex (struct graph *rdg, int v) |
dea61d92 | 995 | { |
526ceb68 | 996 | partition *partition = partition_alloc (NULL, NULL); |
00f96dc9 | 997 | auto_vec<int, 3> nodes; |
24f161fd | 998 | unsigned i; |
dea61d92 SP |
999 | int x; |
1000 | ||
174ec470 | 1001 | graphds_dfs (rdg, &v, 1, &nodes, false, NULL); |
dea61d92 | 1002 | |
9771b263 | 1003 | FOR_EACH_VEC_ELT (nodes, i, x) |
24f161fd RB |
1004 | { |
1005 | bitmap_set_bit (partition->stmts, x); | |
1006 | bitmap_set_bit (partition->loops, | |
1007 | loop_containing_stmt (RDG_STMT (rdg, x))->num); | |
1008 | } | |
dea61d92 | 1009 | |
dea61d92 SP |
1010 | return partition; |
1011 | } | |
1012 | ||
30d55936 RG |
1013 | /* Classifies the builtin kind we can generate for PARTITION of RDG and LOOP. |
1014 | For the moment we detect only the memset zero pattern. */ | |
cfee318d | 1015 | |
30d55936 | 1016 | static void |
526ceb68 | 1017 | classify_partition (loop_p loop, struct graph *rdg, partition *partition) |
cfee318d | 1018 | { |
cfee318d | 1019 | bitmap_iterator bi; |
30d55936 RG |
1020 | unsigned i; |
1021 | tree nb_iter; | |
d0582dc1 | 1022 | data_reference_p single_load, single_store; |
b9fc0497 | 1023 | bool volatiles_p = false; |
d995e887 | 1024 | bool plus_one = false; |
30d55936 RG |
1025 | |
1026 | partition->kind = PKIND_NORMAL; | |
d0582dc1 RG |
1027 | partition->main_dr = NULL; |
1028 | partition->secondary_dr = NULL; | |
818625cf | 1029 | partition->niter = NULL_TREE; |
d995e887 | 1030 | partition->plus_one = false; |
30d55936 | 1031 | |
c61f8985 | 1032 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) |
30d55936 | 1033 | { |
355fe088 | 1034 | gimple *stmt = RDG_STMT (rdg, i); |
30d55936 RG |
1035 | |
1036 | if (gimple_has_volatile_ops (stmt)) | |
b9fc0497 | 1037 | volatiles_p = true; |
cfee318d | 1038 | |
826a536d | 1039 | /* If the stmt has uses outside of the loop mark it as reduction. */ |
9ca86fc3 | 1040 | if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) |
30d55936 | 1041 | { |
826a536d | 1042 | partition->reduction_p = true; |
30d55936 RG |
1043 | return; |
1044 | } | |
1045 | } | |
1046 | ||
b9fc0497 RB |
1047 | /* Perform general partition disqualification for builtins. */ |
1048 | if (volatiles_p | |
1049 | || !flag_tree_loop_distribute_patterns) | |
1050 | return; | |
1051 | ||
d0582dc1 RG |
1052 | /* Detect memset and memcpy. */ |
1053 | single_load = NULL; | |
1054 | single_store = NULL; | |
30d55936 RG |
1055 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) |
1056 | { | |
355fe088 | 1057 | gimple *stmt = RDG_STMT (rdg, i); |
d0582dc1 RG |
1058 | data_reference_p dr; |
1059 | unsigned j; | |
30d55936 RG |
1060 | |
1061 | if (gimple_code (stmt) == GIMPLE_PHI) | |
1062 | continue; | |
1063 | ||
1064 | /* Any scalar stmts are ok. */ | |
1065 | if (!gimple_vuse (stmt)) | |
1066 | continue; | |
1067 | ||
d0582dc1 RG |
1068 | /* Otherwise just regular loads/stores. */ |
1069 | if (!gimple_assign_single_p (stmt)) | |
1070 | return; | |
1071 | ||
1072 | /* But exactly one store and/or load. */ | |
9771b263 | 1073 | for (j = 0; RDG_DATAREFS (rdg, i).iterate (j, &dr); ++j) |
30d55936 | 1074 | { |
d002d099 JJ |
1075 | tree type = TREE_TYPE (DR_REF (dr)); |
1076 | ||
1077 | /* The memset, memcpy and memmove library calls are only | |
1078 | able to deal with generic address space. */ | |
1079 | if (!ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (type))) | |
1080 | return; | |
1081 | ||
d0582dc1 RG |
1082 | if (DR_IS_READ (dr)) |
1083 | { | |
1084 | if (single_load != NULL) | |
1085 | return; | |
1086 | single_load = dr; | |
1087 | } | |
1088 | else | |
1089 | { | |
1090 | if (single_store != NULL) | |
1091 | return; | |
1092 | single_store = dr; | |
1093 | } | |
30d55936 | 1094 | } |
30d55936 RG |
1095 | } |
1096 | ||
818625cf RB |
1097 | if (!single_store) |
1098 | return; | |
1099 | ||
d995e887 | 1100 | nb_iter = number_of_latch_executions (loop); |
818625cf RB |
1101 | if (!nb_iter || nb_iter == chrec_dont_know) |
1102 | return; | |
d995e887 RB |
1103 | if (dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, |
1104 | gimple_bb (DR_STMT (single_store)))) | |
1105 | plus_one = true; | |
818625cf | 1106 | |
d0582dc1 RG |
1107 | if (single_store && !single_load) |
1108 | { | |
355fe088 | 1109 | gimple *stmt = DR_STMT (single_store); |
d0582dc1 | 1110 | tree rhs = gimple_assign_rhs1 (stmt); |
401f3a81 JJ |
1111 | if (const_with_all_bytes_same (rhs) == -1 |
1112 | && (!INTEGRAL_TYPE_P (TREE_TYPE (rhs)) | |
1113 | || (TYPE_MODE (TREE_TYPE (rhs)) | |
1114 | != TYPE_MODE (unsigned_char_type_node)))) | |
d0582dc1 RG |
1115 | return; |
1116 | if (TREE_CODE (rhs) == SSA_NAME | |
1117 | && !SSA_NAME_IS_DEFAULT_DEF (rhs) | |
1118 | && flow_bb_inside_loop_p (loop, gimple_bb (SSA_NAME_DEF_STMT (rhs)))) | |
1119 | return; | |
ca406576 RB |
1120 | if (!adjacent_dr_p (single_store) |
1121 | || !dominated_by_p (CDI_DOMINATORS, | |
1122 | loop->latch, gimple_bb (stmt))) | |
d0582dc1 RG |
1123 | return; |
1124 | partition->kind = PKIND_MEMSET; | |
1125 | partition->main_dr = single_store; | |
818625cf | 1126 | partition->niter = nb_iter; |
d995e887 | 1127 | partition->plus_one = plus_one; |
d0582dc1 RG |
1128 | } |
1129 | else if (single_store && single_load) | |
1130 | { | |
355fe088 TS |
1131 | gimple *store = DR_STMT (single_store); |
1132 | gimple *load = DR_STMT (single_load); | |
d0582dc1 RG |
1133 | /* Direct aggregate copy or via an SSA name temporary. */ |
1134 | if (load != store | |
1135 | && gimple_assign_lhs (load) != gimple_assign_rhs1 (store)) | |
1136 | return; | |
1137 | if (!adjacent_dr_p (single_store) | |
1138 | || !adjacent_dr_p (single_load) | |
1139 | || !operand_equal_p (DR_STEP (single_store), | |
ca406576 RB |
1140 | DR_STEP (single_load), 0) |
1141 | || !dominated_by_p (CDI_DOMINATORS, | |
1142 | loop->latch, gimple_bb (store))) | |
d0582dc1 | 1143 | return; |
f20132e7 RG |
1144 | /* Now check that if there is a dependence this dependence is |
1145 | of a suitable form for memmove. */ | |
6e1aa848 | 1146 | vec<loop_p> loops = vNULL; |
f20132e7 | 1147 | ddr_p ddr; |
9771b263 | 1148 | loops.safe_push (loop); |
f20132e7 RG |
1149 | ddr = initialize_data_dependence_relation (single_load, single_store, |
1150 | loops); | |
1151 | compute_affine_dependence (ddr, loop); | |
f20132e7 RG |
1152 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
1153 | { | |
1154 | free_dependence_relation (ddr); | |
9771b263 | 1155 | loops.release (); |
f20132e7 RG |
1156 | return; |
1157 | } | |
1158 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known) | |
1159 | { | |
1160 | if (DDR_NUM_DIST_VECTS (ddr) == 0) | |
1161 | { | |
1162 | free_dependence_relation (ddr); | |
9771b263 | 1163 | loops.release (); |
f20132e7 RG |
1164 | return; |
1165 | } | |
1166 | lambda_vector dist_v; | |
9771b263 | 1167 | FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) |
f20132e7 RG |
1168 | { |
1169 | int dist = dist_v[index_in_loop_nest (loop->num, | |
1170 | DDR_LOOP_NEST (ddr))]; | |
1171 | if (dist > 0 && !DDR_REVERSED_P (ddr)) | |
1172 | { | |
1173 | free_dependence_relation (ddr); | |
9771b263 | 1174 | loops.release (); |
f20132e7 RG |
1175 | return; |
1176 | } | |
1177 | } | |
510d73a0 | 1178 | partition->kind = PKIND_MEMMOVE; |
f20132e7 | 1179 | } |
510d73a0 RB |
1180 | else |
1181 | partition->kind = PKIND_MEMCPY; | |
b7ce70b3 | 1182 | free_dependence_relation (ddr); |
9771b263 | 1183 | loops.release (); |
d0582dc1 RG |
1184 | partition->main_dr = single_store; |
1185 | partition->secondary_dr = single_load; | |
818625cf | 1186 | partition->niter = nb_iter; |
d995e887 | 1187 | partition->plus_one = plus_one; |
d0582dc1 | 1188 | } |
cfee318d SP |
1189 | } |
1190 | ||
1fa0c180 RG |
1191 | /* For a data reference REF, return the declaration of its base |
1192 | address or NULL_TREE if the base is not determined. */ | |
1193 | ||
1194 | static tree | |
1195 | ref_base_address (data_reference_p dr) | |
1196 | { | |
1197 | tree base_address = DR_BASE_ADDRESS (dr); | |
1198 | if (base_address | |
1199 | && TREE_CODE (base_address) == ADDR_EXPR) | |
1200 | return TREE_OPERAND (base_address, 0); | |
1201 | ||
1202 | return base_address; | |
1203 | } | |
1204 | ||
cfee318d SP |
1205 | /* Returns true when PARTITION1 and PARTITION2 have similar memory |
1206 | accesses in RDG. */ | |
1207 | ||
1208 | static bool | |
526ceb68 TS |
1209 | similar_memory_accesses (struct graph *rdg, partition *partition1, |
1210 | partition *partition2) | |
cfee318d | 1211 | { |
1fa0c180 | 1212 | unsigned i, j, k, l; |
cfee318d | 1213 | bitmap_iterator bi, bj; |
1fa0c180 RG |
1214 | data_reference_p ref1, ref2; |
1215 | ||
1216 | /* First check whether in the intersection of the two partitions are | |
1217 | any loads or stores. Common loads are the situation that happens | |
1218 | most often. */ | |
1219 | EXECUTE_IF_AND_IN_BITMAP (partition1->stmts, partition2->stmts, 0, i, bi) | |
1220 | if (RDG_MEM_WRITE_STMT (rdg, i) | |
1221 | || RDG_MEM_READS_STMT (rdg, i)) | |
1222 | return true; | |
cfee318d | 1223 | |
1fa0c180 | 1224 | /* Then check all data-references against each other. */ |
c61f8985 | 1225 | EXECUTE_IF_SET_IN_BITMAP (partition1->stmts, 0, i, bi) |
cfee318d SP |
1226 | if (RDG_MEM_WRITE_STMT (rdg, i) |
1227 | || RDG_MEM_READS_STMT (rdg, i)) | |
c61f8985 | 1228 | EXECUTE_IF_SET_IN_BITMAP (partition2->stmts, 0, j, bj) |
cfee318d SP |
1229 | if (RDG_MEM_WRITE_STMT (rdg, j) |
1230 | || RDG_MEM_READS_STMT (rdg, j)) | |
1fa0c180 | 1231 | { |
9771b263 | 1232 | FOR_EACH_VEC_ELT (RDG_DATAREFS (rdg, i), k, ref1) |
1fa0c180 RG |
1233 | { |
1234 | tree base1 = ref_base_address (ref1); | |
1235 | if (base1) | |
9771b263 | 1236 | FOR_EACH_VEC_ELT (RDG_DATAREFS (rdg, j), l, ref2) |
1fa0c180 RG |
1237 | if (base1 == ref_base_address (ref2)) |
1238 | return true; | |
1239 | } | |
1240 | } | |
cfee318d SP |
1241 | |
1242 | return false; | |
1243 | } | |
1244 | ||
dea61d92 SP |
1245 | /* Aggregate several components into a useful partition that is |
1246 | registered in the PARTITIONS vector. Partitions will be | |
1247 | distributed in different loops. */ | |
1248 | ||
1249 | static void | |
83a95546 | 1250 | rdg_build_partitions (struct graph *rdg, |
355fe088 | 1251 | vec<gimple *> starting_stmts, |
526ceb68 | 1252 | vec<partition *> *partitions) |
dea61d92 | 1253 | { |
0e3de1d4 | 1254 | auto_bitmap processed; |
2fd5894f | 1255 | int i; |
355fe088 | 1256 | gimple *stmt; |
dea61d92 | 1257 | |
2fd5894f | 1258 | FOR_EACH_VEC_ELT (starting_stmts, i, stmt) |
dea61d92 | 1259 | { |
2fd5894f RB |
1260 | int v = rdg_vertex_for_stmt (rdg, stmt); |
1261 | ||
1262 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1263 | fprintf (dump_file, | |
1264 | "ldist asked to generate code for vertex %d\n", v); | |
b8698a0f | 1265 | |
24f161fd RB |
1266 | /* If the vertex is already contained in another partition so |
1267 | is the partition rooted at it. */ | |
dea61d92 SP |
1268 | if (bitmap_bit_p (processed, v)) |
1269 | continue; | |
b8698a0f | 1270 | |
526ceb68 | 1271 | partition *partition = build_rdg_partition_for_vertex (rdg, v); |
24f161fd | 1272 | bitmap_ior_into (processed, partition->stmts); |
dea61d92 | 1273 | |
826a536d | 1274 | if (dump_file && (dump_flags & TDF_DETAILS)) |
dea61d92 | 1275 | { |
826a536d RB |
1276 | fprintf (dump_file, "ldist useful partition:\n"); |
1277 | dump_bitmap (dump_file, partition->stmts); | |
dea61d92 | 1278 | } |
826a536d RB |
1279 | |
1280 | partitions->safe_push (partition); | |
dea61d92 SP |
1281 | } |
1282 | ||
83a95546 RB |
1283 | /* All vertices should have been assigned to at least one partition now, |
1284 | other than vertices belonging to dead code. */ | |
dea61d92 SP |
1285 | } |
1286 | ||
1287 | /* Dump to FILE the PARTITIONS. */ | |
1288 | ||
1289 | static void | |
526ceb68 | 1290 | dump_rdg_partitions (FILE *file, vec<partition *> partitions) |
dea61d92 SP |
1291 | { |
1292 | int i; | |
526ceb68 | 1293 | partition *partition; |
dea61d92 | 1294 | |
9771b263 | 1295 | FOR_EACH_VEC_ELT (partitions, i, partition) |
c61f8985 | 1296 | debug_bitmap_file (file, partition->stmts); |
dea61d92 SP |
1297 | } |
1298 | ||
1299 | /* Debug PARTITIONS. */ | |
526ceb68 | 1300 | extern void debug_rdg_partitions (vec<partition *> ); |
dea61d92 | 1301 | |
24e47c76 | 1302 | DEBUG_FUNCTION void |
526ceb68 | 1303 | debug_rdg_partitions (vec<partition *> partitions) |
dea61d92 SP |
1304 | { |
1305 | dump_rdg_partitions (stderr, partitions); | |
1306 | } | |
1307 | ||
2b8aee8e SP |
1308 | /* Returns the number of read and write operations in the RDG. */ |
1309 | ||
1310 | static int | |
1311 | number_of_rw_in_rdg (struct graph *rdg) | |
1312 | { | |
1313 | int i, res = 0; | |
1314 | ||
1315 | for (i = 0; i < rdg->n_vertices; i++) | |
1316 | { | |
1317 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1318 | ++res; | |
1319 | ||
1320 | if (RDG_MEM_READS_STMT (rdg, i)) | |
1321 | ++res; | |
1322 | } | |
1323 | ||
1324 | return res; | |
1325 | } | |
1326 | ||
1327 | /* Returns the number of read and write operations in a PARTITION of | |
1328 | the RDG. */ | |
1329 | ||
1330 | static int | |
526ceb68 | 1331 | number_of_rw_in_partition (struct graph *rdg, partition *partition) |
2b8aee8e SP |
1332 | { |
1333 | int res = 0; | |
1334 | unsigned i; | |
1335 | bitmap_iterator ii; | |
1336 | ||
c61f8985 | 1337 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, ii) |
2b8aee8e SP |
1338 | { |
1339 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1340 | ++res; | |
1341 | ||
1342 | if (RDG_MEM_READS_STMT (rdg, i)) | |
1343 | ++res; | |
1344 | } | |
1345 | ||
1346 | return res; | |
1347 | } | |
1348 | ||
1349 | /* Returns true when one of the PARTITIONS contains all the read or | |
1350 | write operations of RDG. */ | |
1351 | ||
1352 | static bool | |
9771b263 | 1353 | partition_contains_all_rw (struct graph *rdg, |
526ceb68 | 1354 | vec<partition *> partitions) |
2b8aee8e SP |
1355 | { |
1356 | int i; | |
526ceb68 | 1357 | partition *partition; |
2b8aee8e SP |
1358 | int nrw = number_of_rw_in_rdg (rdg); |
1359 | ||
9771b263 | 1360 | FOR_EACH_VEC_ELT (partitions, i, partition) |
2b8aee8e SP |
1361 | if (nrw == number_of_rw_in_partition (rdg, partition)) |
1362 | return true; | |
1363 | ||
1364 | return false; | |
1365 | } | |
1366 | ||
447f3223 RB |
1367 | /* Compute partition dependence created by the data references in DRS1 |
1368 | and DRS2 and modify and return DIR according to that. */ | |
1369 | ||
1370 | static int | |
1371 | pg_add_dependence_edges (struct graph *rdg, vec<loop_p> loops, int dir, | |
1372 | vec<data_reference_p> drs1, | |
1373 | vec<data_reference_p> drs2) | |
1374 | { | |
1375 | data_reference_p dr1, dr2; | |
1376 | ||
1377 | /* dependence direction - 0 is no dependence, -1 is back, | |
1378 | 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
1379 | for (int ii = 0; drs1.iterate (ii, &dr1); ++ii) | |
1380 | for (int jj = 0; drs2.iterate (jj, &dr2); ++jj) | |
1381 | { | |
62e22fcb | 1382 | data_reference_p saved_dr1 = dr1; |
2cf19e26 | 1383 | int this_dir = 1; |
447f3223 RB |
1384 | ddr_p ddr; |
1385 | /* Re-shuffle data-refs to be in dominator order. */ | |
1386 | if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
1387 | > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1388 | { | |
6b4db501 | 1389 | std::swap (dr1, dr2); |
447f3223 RB |
1390 | this_dir = -this_dir; |
1391 | } | |
1392 | ddr = initialize_data_dependence_relation (dr1, dr2, loops); | |
1393 | compute_affine_dependence (ddr, loops[0]); | |
1394 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
1395 | this_dir = 2; | |
1396 | else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) | |
1397 | { | |
1398 | if (DDR_REVERSED_P (ddr)) | |
1399 | { | |
6b4db501 | 1400 | std::swap (dr1, dr2); |
447f3223 RB |
1401 | this_dir = -this_dir; |
1402 | } | |
1403 | /* Known dependences can still be unordered througout the | |
1404 | iteration space, see gcc.dg/tree-ssa/ldist-16.c. */ | |
2cf19e26 | 1405 | if (DDR_NUM_DIST_VECTS (ddr) != 1) |
447f3223 | 1406 | this_dir = 2; |
2cf19e26 RB |
1407 | /* If the overlap is exact preserve stmt order. */ |
1408 | else if (lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), 1)) | |
1409 | ; | |
1410 | else | |
1411 | { | |
1412 | /* Else as the distance vector is lexicographic positive | |
1413 | swap the dependence direction. */ | |
1414 | this_dir = -this_dir; | |
1415 | } | |
447f3223 RB |
1416 | } |
1417 | else | |
1418 | this_dir = 0; | |
1419 | free_dependence_relation (ddr); | |
dd18c8c3 JW |
1420 | if (this_dir == 2) |
1421 | return 2; | |
1422 | else if (dir == 0) | |
447f3223 | 1423 | dir = this_dir; |
dd18c8c3 | 1424 | else if (this_dir != 0 && dir != this_dir) |
447f3223 | 1425 | return 2; |
62e22fcb RB |
1426 | /* Shuffle "back" dr1. */ |
1427 | dr1 = saved_dr1; | |
447f3223 RB |
1428 | } |
1429 | return dir; | |
1430 | } | |
1431 | ||
1432 | /* Compare postorder number of the partition graph vertices V1 and V2. */ | |
1433 | ||
1434 | static int | |
1435 | pgcmp (const void *v1_, const void *v2_) | |
1436 | { | |
1437 | const vertex *v1 = (const vertex *)v1_; | |
1438 | const vertex *v2 = (const vertex *)v2_; | |
1439 | return v2->post - v1->post; | |
1440 | } | |
2fd5894f RB |
1441 | |
1442 | /* Distributes the code from LOOP in such a way that producer | |
1443 | statements are placed before consumer statements. Tries to separate | |
1444 | only the statements from STMTS into separate loops. | |
b71b7a8e RB |
1445 | Returns the number of distributed loops. Set *DESTROY_P to whether |
1446 | LOOP needs to be destroyed. */ | |
dea61d92 SP |
1447 | |
1448 | static int | |
355fe088 | 1449 | distribute_loop (struct loop *loop, vec<gimple *> stmts, |
b71b7a8e | 1450 | control_dependences *cd, int *nb_calls, bool *destroy_p) |
dea61d92 | 1451 | { |
2fd5894f | 1452 | struct graph *rdg; |
526ceb68 | 1453 | partition *partition; |
be6b029b | 1454 | bool any_builtin; |
2fd5894f | 1455 | int i, nbp; |
447f3223 RB |
1456 | graph *pg = NULL; |
1457 | int num_sccs = 1; | |
dea61d92 | 1458 | |
c9326aef | 1459 | *destroy_p = false; |
826a536d | 1460 | *nb_calls = 0; |
00f96dc9 | 1461 | auto_vec<loop_p, 3> loop_nest; |
2fd5894f | 1462 | if (!find_loop_nest (loop, &loop_nest)) |
07687835 | 1463 | return 0; |
2fd5894f | 1464 | |
36875e8f | 1465 | rdg = build_rdg (loop_nest, cd); |
2fd5894f RB |
1466 | if (!rdg) |
1467 | { | |
1468 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1469 | fprintf (dump_file, | |
1470 | "Loop %d not distributed: failed to build the RDG.\n", | |
1471 | loop->num); | |
1472 | ||
2fd5894f RB |
1473 | return 0; |
1474 | } | |
1475 | ||
1476 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1477 | dump_rdg (dump_file, rdg); | |
1478 | ||
526ceb68 | 1479 | auto_vec<struct partition *, 3> partitions; |
2fd5894f | 1480 | rdg_build_partitions (rdg, stmts, &partitions); |
dea61d92 | 1481 | |
be6b029b | 1482 | any_builtin = false; |
9771b263 | 1483 | FOR_EACH_VEC_ELT (partitions, i, partition) |
be6b029b RG |
1484 | { |
1485 | classify_partition (loop, rdg, partition); | |
1486 | any_builtin |= partition_builtin_p (partition); | |
1487 | } | |
30d55936 | 1488 | |
447f3223 RB |
1489 | /* If we are only distributing patterns but did not detect any, |
1490 | simply bail out. */ | |
9fed7f3a RB |
1491 | if (!flag_tree_loop_distribution |
1492 | && !any_builtin) | |
1493 | { | |
1494 | nbp = 0; | |
1495 | goto ldist_done; | |
1496 | } | |
1497 | ||
447f3223 RB |
1498 | /* If we are only distributing patterns fuse all partitions that |
1499 | were not classified as builtins. This also avoids chopping | |
1500 | a loop into pieces, separated by builtin calls. That is, we | |
1501 | only want no or a single loop body remaining. */ | |
526ceb68 | 1502 | struct partition *into; |
447f3223 RB |
1503 | if (!flag_tree_loop_distribution) |
1504 | { | |
1505 | for (i = 0; partitions.iterate (i, &into); ++i) | |
1506 | if (!partition_builtin_p (into)) | |
1507 | break; | |
1508 | for (++i; partitions.iterate (i, &partition); ++i) | |
1509 | if (!partition_builtin_p (partition)) | |
1510 | { | |
1511 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1512 | { | |
1513 | fprintf (dump_file, "fusing non-builtin partitions\n"); | |
1514 | dump_bitmap (dump_file, into->stmts); | |
1515 | dump_bitmap (dump_file, partition->stmts); | |
1516 | } | |
1517 | partition_merge_into (into, partition); | |
1518 | partitions.unordered_remove (i); | |
1519 | partition_free (partition); | |
1520 | i--; | |
1521 | } | |
1522 | } | |
1523 | ||
1524 | /* Due to limitations in the transform phase we have to fuse all | |
1525 | reduction partitions into the last partition so the existing | |
1526 | loop will contain all loop-closed PHI nodes. */ | |
1527 | for (i = 0; partitions.iterate (i, &into); ++i) | |
1528 | if (partition_reduction_p (into)) | |
1529 | break; | |
1530 | for (i = i + 1; partitions.iterate (i, &partition); ++i) | |
1531 | if (partition_reduction_p (partition)) | |
1532 | { | |
1533 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1534 | { | |
1535 | fprintf (dump_file, "fusing partitions\n"); | |
1536 | dump_bitmap (dump_file, into->stmts); | |
1537 | dump_bitmap (dump_file, partition->stmts); | |
1538 | fprintf (dump_file, "because they have reductions\n"); | |
1539 | } | |
1540 | partition_merge_into (into, partition); | |
1541 | partitions.unordered_remove (i); | |
1542 | partition_free (partition); | |
1543 | i--; | |
1544 | } | |
1545 | ||
9fed7f3a RB |
1546 | /* Apply our simple cost model - fuse partitions with similar |
1547 | memory accesses. */ | |
9fed7f3a RB |
1548 | for (i = 0; partitions.iterate (i, &into); ++i) |
1549 | { | |
16eba420 | 1550 | bool changed = false; |
9fed7f3a RB |
1551 | if (partition_builtin_p (into)) |
1552 | continue; | |
1553 | for (int j = i + 1; | |
1554 | partitions.iterate (j, &partition); ++j) | |
1555 | { | |
40b6bff9 | 1556 | if (similar_memory_accesses (rdg, into, partition)) |
9fed7f3a RB |
1557 | { |
1558 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1559 | { | |
1560 | fprintf (dump_file, "fusing partitions\n"); | |
1561 | dump_bitmap (dump_file, into->stmts); | |
1562 | dump_bitmap (dump_file, partition->stmts); | |
1563 | fprintf (dump_file, "because they have similar " | |
1564 | "memory accesses\n"); | |
1565 | } | |
826a536d | 1566 | partition_merge_into (into, partition); |
447f3223 | 1567 | partitions.unordered_remove (j); |
9fed7f3a RB |
1568 | partition_free (partition); |
1569 | j--; | |
16eba420 | 1570 | changed = true; |
9fed7f3a RB |
1571 | } |
1572 | } | |
16eba420 RB |
1573 | /* If we fused 0 1 2 in step 1 to 0,2 1 as 0 and 2 have similar |
1574 | accesses when 1 and 2 have similar accesses but not 0 and 1 | |
1575 | then in the next iteration we will fail to consider merging | |
1576 | 1 into 0,2. So try again if we did any merging into 0. */ | |
1577 | if (changed) | |
1578 | i--; | |
9fed7f3a RB |
1579 | } |
1580 | ||
447f3223 RB |
1581 | /* Build the partition dependency graph. */ |
1582 | if (partitions.length () > 1) | |
c014f6f5 | 1583 | { |
447f3223 RB |
1584 | pg = new_graph (partitions.length ()); |
1585 | struct pgdata { | |
526ceb68 | 1586 | struct partition *partition; |
447f3223 RB |
1587 | vec<data_reference_p> writes; |
1588 | vec<data_reference_p> reads; | |
1589 | }; | |
1590 | #define PGDATA(i) ((pgdata *)(pg->vertices[i].data)) | |
1591 | for (i = 0; partitions.iterate (i, &partition); ++i) | |
1592 | { | |
1593 | vertex *v = &pg->vertices[i]; | |
1594 | pgdata *data = new pgdata; | |
1595 | data_reference_p dr; | |
1596 | /* FIXME - leaks. */ | |
1597 | v->data = data; | |
1598 | bitmap_iterator bi; | |
1599 | unsigned j; | |
1600 | data->partition = partition; | |
1601 | data->reads = vNULL; | |
1602 | data->writes = vNULL; | |
1603 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, j, bi) | |
1604 | for (int k = 0; RDG_DATAREFS (rdg, j).iterate (k, &dr); ++k) | |
1605 | if (DR_IS_READ (dr)) | |
1606 | data->reads.safe_push (dr); | |
1607 | else | |
1608 | data->writes.safe_push (dr); | |
1609 | } | |
526ceb68 | 1610 | struct partition *partition1, *partition2; |
447f3223 RB |
1611 | for (i = 0; partitions.iterate (i, &partition1); ++i) |
1612 | for (int j = i + 1; partitions.iterate (j, &partition2); ++j) | |
1613 | { | |
1614 | /* dependence direction - 0 is no dependence, -1 is back, | |
1615 | 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
1616 | int dir = 0; | |
1617 | dir = pg_add_dependence_edges (rdg, loop_nest, dir, | |
1618 | PGDATA(i)->writes, | |
1619 | PGDATA(j)->reads); | |
1620 | if (dir != 2) | |
1621 | dir = pg_add_dependence_edges (rdg, loop_nest, dir, | |
1622 | PGDATA(i)->reads, | |
1623 | PGDATA(j)->writes); | |
1624 | if (dir != 2) | |
1625 | dir = pg_add_dependence_edges (rdg, loop_nest, dir, | |
1626 | PGDATA(i)->writes, | |
1627 | PGDATA(j)->writes); | |
1628 | if (dir == 1 || dir == 2) | |
1629 | add_edge (pg, i, j); | |
1630 | if (dir == -1 || dir == 2) | |
1631 | add_edge (pg, j, i); | |
1632 | } | |
1633 | ||
1634 | /* Add edges to the reduction partition (if any) to force it last. */ | |
1635 | unsigned j; | |
1636 | for (j = 0; partitions.iterate (j, &partition); ++j) | |
1637 | if (partition_reduction_p (partition)) | |
1638 | break; | |
1639 | if (j < partitions.length ()) | |
38ad2d07 | 1640 | { |
447f3223 RB |
1641 | for (unsigned i = 0; partitions.iterate (i, &partition); ++i) |
1642 | if (i != j) | |
1643 | add_edge (pg, i, j); | |
1644 | } | |
1645 | ||
1646 | /* Compute partitions we cannot separate and fuse them. */ | |
1647 | num_sccs = graphds_scc (pg, NULL); | |
1648 | for (i = 0; i < num_sccs; ++i) | |
1649 | { | |
526ceb68 | 1650 | struct partition *first; |
447f3223 RB |
1651 | int j; |
1652 | for (j = 0; partitions.iterate (j, &first); ++j) | |
1653 | if (pg->vertices[j].component == i) | |
38ad2d07 | 1654 | break; |
447f3223 RB |
1655 | for (j = j + 1; partitions.iterate (j, &partition); ++j) |
1656 | if (pg->vertices[j].component == i) | |
38ad2d07 | 1657 | { |
447f3223 RB |
1658 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1659 | { | |
1660 | fprintf (dump_file, "fusing partitions\n"); | |
1661 | dump_bitmap (dump_file, first->stmts); | |
1662 | dump_bitmap (dump_file, partition->stmts); | |
1663 | fprintf (dump_file, "because they are in the same " | |
1664 | "dependence SCC\n"); | |
1665 | } | |
1666 | partition_merge_into (first, partition); | |
1667 | partitions[j] = NULL; | |
5eb010bc | 1668 | partition_free (partition); |
447f3223 | 1669 | PGDATA (j)->partition = NULL; |
38ad2d07 | 1670 | } |
38ad2d07 | 1671 | } |
30d55936 | 1672 | |
447f3223 RB |
1673 | /* Now order the remaining nodes in postorder. */ |
1674 | qsort (pg->vertices, pg->n_vertices, sizeof (vertex), pgcmp); | |
1675 | partitions.truncate (0); | |
1676 | for (i = 0; i < pg->n_vertices; ++i) | |
b9fc0497 | 1677 | { |
447f3223 RB |
1678 | pgdata *data = PGDATA (i); |
1679 | if (data->partition) | |
1680 | partitions.safe_push (data->partition); | |
1681 | data->reads.release (); | |
1682 | data->writes.release (); | |
1683 | delete data; | |
b9fc0497 | 1684 | } |
447f3223 RB |
1685 | gcc_assert (partitions.length () == (unsigned)num_sccs); |
1686 | free_graph (pg); | |
b9fc0497 RB |
1687 | } |
1688 | ||
9771b263 | 1689 | nbp = partitions.length (); |
a4293fa6 | 1690 | if (nbp == 0 |
9771b263 DN |
1691 | || (nbp == 1 && !partition_builtin_p (partitions[0])) |
1692 | || (nbp > 1 && partition_contains_all_rw (rdg, partitions))) | |
c014f6f5 RG |
1693 | { |
1694 | nbp = 0; | |
1695 | goto ldist_done; | |
1696 | } | |
dea61d92 SP |
1697 | |
1698 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1699 | dump_rdg_partitions (dump_file, partitions); | |
1700 | ||
9771b263 | 1701 | FOR_EACH_VEC_ELT (partitions, i, partition) |
826a536d RB |
1702 | { |
1703 | if (partition_builtin_p (partition)) | |
1704 | (*nb_calls)++; | |
b71b7a8e | 1705 | *destroy_p |= generate_code_for_partition (loop, partition, i < nbp - 1); |
826a536d | 1706 | } |
dea61d92 | 1707 | |
dea61d92 SP |
1708 | ldist_done: |
1709 | ||
9771b263 | 1710 | FOR_EACH_VEC_ELT (partitions, i, partition) |
c61f8985 | 1711 | partition_free (partition); |
dea61d92 | 1712 | |
dea61d92 | 1713 | free_rdg (rdg); |
826a536d | 1714 | return nbp - *nb_calls; |
dea61d92 SP |
1715 | } |
1716 | ||
1717 | /* Distribute all loops in the current function. */ | |
1718 | ||
be55bfe6 TS |
1719 | namespace { |
1720 | ||
1721 | const pass_data pass_data_loop_distribution = | |
1722 | { | |
1723 | GIMPLE_PASS, /* type */ | |
1724 | "ldist", /* name */ | |
1725 | OPTGROUP_LOOP, /* optinfo_flags */ | |
be55bfe6 TS |
1726 | TV_TREE_LOOP_DISTRIBUTION, /* tv_id */ |
1727 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
1728 | 0, /* properties_provided */ | |
1729 | 0, /* properties_destroyed */ | |
1730 | 0, /* todo_flags_start */ | |
3bea341f | 1731 | 0, /* todo_flags_finish */ |
be55bfe6 TS |
1732 | }; |
1733 | ||
1734 | class pass_loop_distribution : public gimple_opt_pass | |
1735 | { | |
1736 | public: | |
1737 | pass_loop_distribution (gcc::context *ctxt) | |
1738 | : gimple_opt_pass (pass_data_loop_distribution, ctxt) | |
1739 | {} | |
1740 | ||
1741 | /* opt_pass methods: */ | |
1742 | virtual bool gate (function *) | |
1743 | { | |
1744 | return flag_tree_loop_distribution | |
1745 | || flag_tree_loop_distribute_patterns; | |
1746 | } | |
1747 | ||
1748 | virtual unsigned int execute (function *); | |
1749 | ||
1750 | }; // class pass_loop_distribution | |
1751 | ||
1752 | unsigned int | |
1753 | pass_loop_distribution::execute (function *fun) | |
dea61d92 SP |
1754 | { |
1755 | struct loop *loop; | |
c014f6f5 | 1756 | bool changed = false; |
1fa0c180 | 1757 | basic_block bb; |
36875e8f | 1758 | control_dependences *cd = NULL; |
b71b7a8e | 1759 | auto_vec<loop_p> loops_to_be_destroyed; |
1fa0c180 | 1760 | |
773d9217 BC |
1761 | if (number_of_loops (fun) <= 1) |
1762 | return 0; | |
1763 | ||
be55bfe6 | 1764 | FOR_ALL_BB_FN (bb, fun) |
1fa0c180 RG |
1765 | { |
1766 | gimple_stmt_iterator gsi; | |
1767 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1768 | gimple_set_uid (gsi_stmt (gsi), -1); | |
1769 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1770 | gimple_set_uid (gsi_stmt (gsi), -1); | |
1771 | } | |
dea61d92 | 1772 | |
c014f6f5 RG |
1773 | /* We can at the moment only distribute non-nested loops, thus restrict |
1774 | walking to innermost loops. */ | |
f0bd40b1 | 1775 | FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) |
dea61d92 | 1776 | { |
355fe088 | 1777 | auto_vec<gimple *> work_list; |
be6b029b | 1778 | basic_block *bbs; |
0e20c89f | 1779 | int num = loop->num; |
be6b029b | 1780 | unsigned int i; |
a3357f7d RG |
1781 | |
1782 | /* If the loop doesn't have a single exit we will fail anyway, | |
1783 | so do that early. */ | |
1784 | if (!single_exit (loop)) | |
1785 | continue; | |
dea61d92 | 1786 | |
f56f2d33 JH |
1787 | /* Only optimize hot loops. */ |
1788 | if (!optimize_loop_for_speed_p (loop)) | |
1789 | continue; | |
1790 | ||
be6b029b RG |
1791 | /* Initialize the worklist with stmts we seed the partitions with. */ |
1792 | bbs = get_loop_body_in_dom_order (loop); | |
1793 | for (i = 0; i < loop->num_nodes; ++i) | |
1794 | { | |
538dd0b7 DM |
1795 | for (gphi_iterator gsi = gsi_start_phis (bbs[i]); |
1796 | !gsi_end_p (gsi); | |
1797 | gsi_next (&gsi)) | |
deb6c11a | 1798 | { |
538dd0b7 | 1799 | gphi *phi = gsi.phi (); |
deb6c11a RB |
1800 | if (virtual_operand_p (gimple_phi_result (phi))) |
1801 | continue; | |
1802 | /* Distribute stmts which have defs that are used outside of | |
be55bfe6 | 1803 | the loop. */ |
deb6c11a RB |
1804 | if (!stmt_has_scalar_dependences_outside_loop (loop, phi)) |
1805 | continue; | |
1806 | work_list.safe_push (phi); | |
1807 | } | |
538dd0b7 DM |
1808 | for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); |
1809 | !gsi_end_p (gsi); | |
1810 | gsi_next (&gsi)) | |
be6b029b | 1811 | { |
355fe088 | 1812 | gimple *stmt = gsi_stmt (gsi); |
83a95546 RB |
1813 | |
1814 | /* If there is a stmt with side-effects bail out - we | |
be55bfe6 | 1815 | cannot and should not distribute this loop. */ |
83a95546 RB |
1816 | if (gimple_has_side_effects (stmt)) |
1817 | { | |
1818 | work_list.truncate (0); | |
1819 | goto out; | |
1820 | } | |
1821 | ||
b9fc0497 | 1822 | /* Distribute stmts which have defs that are used outside of |
be55bfe6 | 1823 | the loop. */ |
b9fc0497 RB |
1824 | if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) |
1825 | ; | |
1826 | /* Otherwise only distribute stores for now. */ | |
e179190c | 1827 | else if (!gimple_vdef (stmt)) |
be6b029b RG |
1828 | continue; |
1829 | ||
9771b263 | 1830 | work_list.safe_push (stmt); |
be6b029b RG |
1831 | } |
1832 | } | |
83a95546 | 1833 | out: |
be6b029b | 1834 | free (bbs); |
c014f6f5 | 1835 | |
826a536d RB |
1836 | int nb_generated_loops = 0; |
1837 | int nb_generated_calls = 0; | |
1838 | location_t loc = find_loop_location (loop); | |
9771b263 | 1839 | if (work_list.length () > 0) |
36875e8f RB |
1840 | { |
1841 | if (!cd) | |
1842 | { | |
ca406576 | 1843 | calculate_dominance_info (CDI_DOMINATORS); |
36875e8f | 1844 | calculate_dominance_info (CDI_POST_DOMINATORS); |
30fd2977 | 1845 | cd = new control_dependences (); |
36875e8f RB |
1846 | free_dominance_info (CDI_POST_DOMINATORS); |
1847 | } | |
b71b7a8e | 1848 | bool destroy_p; |
826a536d | 1849 | nb_generated_loops = distribute_loop (loop, work_list, cd, |
b71b7a8e RB |
1850 | &nb_generated_calls, |
1851 | &destroy_p); | |
1852 | if (destroy_p) | |
1853 | loops_to_be_destroyed.safe_push (loop); | |
36875e8f | 1854 | } |
c014f6f5 | 1855 | |
826a536d | 1856 | if (nb_generated_loops + nb_generated_calls > 0) |
dea61d92 | 1857 | { |
826a536d RB |
1858 | changed = true; |
1859 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, | |
1860 | loc, "Loop %d distributed: split to %d loops " | |
1861 | "and %d library calls.\n", | |
1862 | num, nb_generated_loops, nb_generated_calls); | |
dea61d92 | 1863 | } |
826a536d RB |
1864 | else if (dump_file && (dump_flags & TDF_DETAILS)) |
1865 | fprintf (dump_file, "Loop %d is the same.\n", num); | |
dea61d92 SP |
1866 | } |
1867 | ||
36875e8f RB |
1868 | if (cd) |
1869 | delete cd; | |
1870 | ||
c014f6f5 RG |
1871 | if (changed) |
1872 | { | |
30fd2977 RB |
1873 | /* Destroy loop bodies that could not be reused. Do this late as we |
1874 | otherwise can end up refering to stale data in control dependences. */ | |
1875 | unsigned i; | |
1876 | FOR_EACH_VEC_ELT (loops_to_be_destroyed, i, loop) | |
1877 | destroy_loop (loop); | |
1878 | ||
d0ed943c RB |
1879 | /* Cached scalar evolutions now may refer to wrong or non-existing |
1880 | loops. */ | |
1881 | scev_reset_htab (); | |
be55bfe6 | 1882 | mark_virtual_operands_for_renaming (fun); |
c014f6f5 RG |
1883 | rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); |
1884 | } | |
1885 | ||
b2b29377 | 1886 | checking_verify_loop_structure (); |
c014f6f5 | 1887 | |
5006671f | 1888 | return 0; |
dea61d92 SP |
1889 | } |
1890 | ||
27a4cd48 DM |
1891 | } // anon namespace |
1892 | ||
1893 | gimple_opt_pass * | |
1894 | make_pass_loop_distribution (gcc::context *ctxt) | |
1895 | { | |
1896 | return new pass_loop_distribution (ctxt); | |
1897 | } |